Calculating weights for Gradient Blending for new datasets

[AntiLibrary5]

Hi,
Are there any additional resources about calculating the normalizing constants for the blending weights from a practical implementation point of view apart from the literature (what makes training multi-modal classification networks hard?).

The weights for the datasets used are given : https://github.com/facebookresearch/VMZ/blob/master/c2/tutorials/gradient_blending.md
But I couldn't find any code where their calculation is implemented.

But more importantly how to find them for a new dataset? The proof of the proposition for gradient blending is not helpful in this regard. Also going through the repo, they are considered an argument:

VMZ/c2/tools/train_net.py

Line 255 in 4cc542d

return model_builder.build_model(

in the "add_weighted_loss()" method and I couldn't find any helper functions which might have been used to calculate them.

Thank you for your response.

[jc5201]

Any changes here?